HEAS: Hierarchical Evolutionary Agent-Based Simulation Framework for Multi-Objective Policy Search

A team of researchers has introduced HEAS (Hierarchical Evolutionary Agent-based Simulation), a new Python framework designed to solve a critical but often overlooked problem in AI agent training: metric aggregation divergence. This occurs when different stages of a simulation pipeline—like optimization, evaluation, and validation—independently calculate performance metrics for AI policies (the rules agents follow). The slight differences in these calculations create a hidden confound, meaning the 'champion' policy selected might be an artifact of this inconsistency rather than genuinely superior. HEAS eliminates this by enforcing a uniform 'metric contract'—a single, shared function for calculating outcomes—across all stages, ensuring apples-to-apples comparisons.

In controlled experiments, the impact is substantial. HEAS reduced rank reversals (where a worse policy incorrectly beats a better one) by 50% compared to ad-hoc methods. In one test, the policy champion selected by HEAS won all 32 held-out ecological scenarios, a result that would be uninterpretable under traditional, divergent systems. Beyond accuracy, the framework dramatically improves developer efficiency. By standardizing metric handling, HEAS reduced the amount of 'coupling code'—the glue connecting different components—by 97%, from 160 lines to just 5 compared to the popular Mesa 3.3.1 library.

The framework's composability has been validated across diverse domains, including ecological systems, enterprise operations, and mean-field ordinary differential equations, proving its utility for multi-objective policy search in complex, simulated environments. The researchers have released HEAS as an open-source Python package alongside a web playground, making this robust tool for building reliable multi-agent AI simulations accessible to the broader community.